16 ms·

> This statement is incredibly biased That is correct. But the question is: Why is that a problem? Biased against exploitation and overwork is good. Biased aga

by sgift 4y ago

> This statement is incredibly biased That is correct. But the question is: Why is that a problem? Biased against exploitation and overwork is good. Biased against accelerating the climate crisis is good. Biased against discrimination is good. I fail to see which of these biases is bad here.

version_five 4y ago
When you clearly indicate you're not neutral, you lose all credibility. Nobody wants a ML model that gives them the climate warrior version of the "truth". Neutrality is extremely important in order to be broadly taken seriously. It's exactly the kind of criticism that's been leveled against chatGPT
- kokanee 4y ago
  Ah yes, I'll just reference the list of my non-neutral biases as I choose the moderation rules for my AI. My bias against swear words is neutral, so I will include that rule, but my bias against pollution is not neutral, so I will skip those moderation rules. Obviously categorizing beliefs into "neutral" and "not neutral" is impossible. Your statement is a classic example of the false consensus effect -- everyone thinks they are neutral. https://en.wikipedia.org/wiki/False_consensus_effect https://en.wikipedia.org/wiki/False_consensus_effect
  - haberman 4y ago
    > Obviously categorizing beliefs into "neutral" and "not neutral" is impossible. There is one distinction that matters a lot, and that is empirical vs. moral. Also known as vs. ought (https://en.wikipedia.org/wiki/Is%E2%80%93ought_problem https://en.wikipedia.org/wiki/Is%E2%80%93ought_problem). When exploring empirical claims, one can aspire for neutrality. Even if perfect neutrality is not possible, it's certainly possible to do better or worse at it. Moral claims cannot really be evaluated neutrally, I agree. But an AI striving for neutrality could attempt to avoid taking sides on contentious issues.
- notahacker 4y ago
  The irony with people talking about "neutrality" is the tendency of the people making such demands to be even more obsessed with distorting the input data to produce outcomes censored to take into account their viewpoint than the 'AI safety' and PR people. I mean, how much censorship (or artificial curation) would you need to avoid an ML model giving "the climate warrior version" of questions about whether the world was getting warmer?!
Avicebron 4y ago
Because there isn't a universal truth ), at least if there is we as a species don't (can't know it) especially as it relates to how we all interact not only with each other but the planet, etc. You're version of good is another's version of bad, if we can't have neutrality, we're just building another machine to amplify whichever group builds it's values and right or wrong just depends on where you stand. To break it down, do we want to be neutral or do we want, SiliconValleyGPT? What happens when instead of that we get SaudiArabiaGPT? Or ChinaGPT? Or RussiaGPT? DeepSouthGPT? I just picked arbitrary places but you see my point, I hope.
- MrJohz 4y ago
  There's an assumption there that a neutral AI can exist, but I think a lot of people would challenge that central assumption. No set of training data can be truly neutral or unbiased. It may be well balanced between certain groups, but the choice of which groups to balance, how much to balance them, and the choice to add this balance in the first place are all ideological decisions that stem from certain beliefs and ideas. The AIs that get built will reflect the ideologies and values of the people who build them. It is therefore better for us ethically to be conscious of the values we are injecting.
- danShumway 4y ago
  These kinds of philosophy discussions are frustratingly restricted to bias against minorities. Nobody here commented on the "AI should protect your privacy" tenant with "but how do we know privacy is good? What if my definition of privacy is different from yours? What happens when a criminal has privacy?" Nobody wanted a concrete definition of agency from first principles, nobody wanted to talk about the intersection of agency and telos. "There's no universal truth" is basically an argument against "responsible" AI in the first place, since there would be no universal truth about what "responsibility" means. Mozilla's statement about responsible AI is inherently biased towards their opinion of what responsibility is. But again, the bias accusations only popped up on that last point. We're all fine with Mozilla having opinions about "good" and "bad" states of the world until it has opinions about treating minorities equitably, then it becomes pressingly important that we have a philosophy discussion.
boringuser2 4y ago
Look at where your failure lies here -- you literally just asserted your biases as "good" and said you failed to see another perspective. That's literally the point.
- sgift 4y ago
  I didn't fail to see another perspective. I rejected the other positions as worse (after careful examination). That's different. Also, neutrality is for the most part just a status quo bias. "Things are good as they are" is a position as much as every other.
  - none_to_remain 4y ago
    Zero distinction between "is" and "ought"
- kokanee 4y ago
  Your point seems to be that AI output should not be moderated. This would mean that the AI would adopt whatever biases and language patterns exist in the training data. In that scenario, the AI developer is still injecting bias by selecting which training data to use. There's also the problem that any unmoderated AI would be completely commercially unviable, of course. So, I think I understand what you're opposed to, but I'm curious what actions/methodologies you would be in favor of.
the_third_wave 4y ago
Because it implies the "climate crisis" is a real thing. For some it surely is, others - me among them - see this differently. Time will tell who got it right but just the fact that the media more or less dropped the "climate" scare when SARS2 hit the front pages should give those who are in the former camp something to think about. Only when it became clear that there was no more to be gained from pushing SARS2 scare stories did they return to the climate narrative. A LLM which has been trained to push "climate crisis" will end up producing agitprop [1] instead of objective output. The same would be true for a model which has been trained to deny anything related to climate change but thus far I have not seen any call for such training methods to be used. [1] https://www.britannica.com/topic/agitprop https://www.britannica.com/topic/agitprop
freehorse 4y ago
But this is exactly the problem: if there was human alignment on these issues, we would have been already on the way of tackling them. But there is not; the planet is still getting destroyed for profit, racial etc discrimination is still ongoing etc. Tackling these would require us to turn down different ethics and human rules already in place like economic liberalism, concept of copyright etc. That would class with other rights etc in place. Human alignment is maintained through several rules and relations comprising a psychological, social and cultural complex, and even then we are witnessing continued ecological collapse, wars, corruption, social unrest etc. What is AI supposed to align with, first of all, before we even figure out the "how" (as AI does not fall in the sphere of those relations described above that affect human alignment)?

by sgift 4y ago

version_five 4y ago

When you clearly indicate you're not neutral, you lose all credibility. Nobody wants a ML model that gives them the climate warrior version of the "truth". Neutrality is extremely important in order to be broadly taken seriously. It's exactly the kind of criticism that's been leveled against chatGPT

kokanee 4y ago
Ah yes, I'll just reference the list of my non-neutral biases as I choose the moderation rules for my AI. My bias against swear words is neutral, so I will include that rule, but my bias against pollution is not neutral, so I will skip those moderation rules. Obviously categorizing beliefs into "neutral" and "not neutral" is impossible. Your statement is a classic example of the false consensus effect -- everyone thinks they are neutral. https://en.wikipedia.org/wiki/False_consensus_effect https://en.wikipedia.org/wiki/False_consensus_effect
- haberman 4y ago
  > Obviously categorizing beliefs into "neutral" and "not neutral" is impossible. There is one distinction that matters a lot, and that is empirical vs. moral. Also known as vs. ought (https://en.wikipedia.org/wiki/Is%E2%80%93ought_problem https://en.wikipedia.org/wiki/Is%E2%80%93ought_problem). When exploring empirical claims, one can aspire for neutrality. Even if perfect neutrality is not possible, it's certainly possible to do better or worse at it. Moral claims cannot really be evaluated neutrally, I agree. But an AI striving for neutrality could attempt to avoid taking sides on contentious issues.
notahacker 4y ago
The irony with people talking about "neutrality" is the tendency of the people making such demands to be even more obsessed with distorting the input data to produce outcomes censored to take into account their viewpoint than the 'AI safety' and PR people. I mean, how much censorship (or artificial curation) would you need to avoid an ML model giving "the climate warrior version" of questions about whether the world was getting warmer?!

Avicebron 4y ago

Because there isn't a universal truth ), at least if there is we as a species don't (can't know it) especially as it relates to how we all interact not only with each other but the planet, etc. You're version of good is another's version of bad, if we can't have neutrality, we're just building another machine to amplify whichever group builds it's values and right or wrong just depends on where you stand. To break it down, do we want to be neutral or do we want, SiliconValleyGPT? What happens when instead of that we get SaudiArabiaGPT? Or ChinaGPT? Or RussiaGPT? DeepSouthGPT? I just picked arbitrary places but you see my point, I hope.

MrJohz 4y ago
There's an assumption there that a neutral AI can exist, but I think a lot of people would challenge that central assumption. No set of training data can be truly neutral or unbiased. It may be well balanced between certain groups, but the choice of which groups to balance, how much to balance them, and the choice to add this balance in the first place are all ideological decisions that stem from certain beliefs and ideas. The AIs that get built will reflect the ideologies and values of the people who build them. It is therefore better for us ethically to be conscious of the values we are injecting.
danShumway 4y ago
These kinds of philosophy discussions are frustratingly restricted to bias against minorities. Nobody here commented on the "AI should protect your privacy" tenant with "but how do we know privacy is good? What if my definition of privacy is different from yours? What happens when a criminal has privacy?" Nobody wanted a concrete definition of agency from first principles, nobody wanted to talk about the intersection of agency and telos. "There's no universal truth" is basically an argument against "responsible" AI in the first place, since there would be no universal truth about what "responsibility" means. Mozilla's statement about responsible AI is inherently biased towards their opinion of what responsibility is. But again, the bias accusations only popped up on that last point. We're all fine with Mozilla having opinions about "good" and "bad" states of the world until it has opinions about treating minorities equitably, then it becomes pressingly important that we have a philosophy discussion.

boringuser2 4y ago

Look at where your failure lies here -- you literally just asserted your biases as "good" and said you failed to see another perspective. That's literally the point.

sgift 4y ago
I didn't fail to see another perspective. I rejected the other positions as worse (after careful examination). That's different. Also, neutrality is for the most part just a status quo bias. "Things are good as they are" is a position as much as every other.
- none_to_remain 4y ago
  Zero distinction between "is" and "ought"
kokanee 4y ago
Your point seems to be that AI output should not be moderated. This would mean that the AI would adopt whatever biases and language patterns exist in the training data. In that scenario, the AI developer is still injecting bias by selecting which training data to use. There's also the problem that any unmoderated AI would be completely commercially unviable, of course. So, I think I understand what you're opposed to, but I'm curious what actions/methodologies you would be in favor of.

the_third_wave 4y ago

Because it implies the "climate crisis" is a real thing. For some it surely is, others - me among them - see this differently. Time will tell who got it right but just the fact that the media more or less dropped the "climate" scare when SARS2 hit the front pages should give those who are in the former camp something to think about. Only when it became clear that there was no more to be gained from pushing SARS2 scare stories did they return to the climate narrative. A LLM which has been trained to push "climate crisis" will end up producing agitprop [1] instead of objective output. The same would be true for a model which has been trained to deny anything related to climate change but thus far I have not seen any call for such training methods to be used. [1] https://www.britannica.com/topic/agitprop https://www.britannica.com/topic/agitprop

freehorse 4y ago

But this is exactly the problem: if there was human alignment on these issues, we would have been already on the way of tackling them. But there is not; the planet is still getting destroyed for profit, racial etc discrimination is still ongoing etc. Tackling these would require us to turn down different ethics and human rules already in place like economic liberalism, concept of copyright etc. That would class with other rights etc in place. Human alignment is maintained through several rules and relations comprising a psychological, social and cultural complex, and even then we are witnessing continued ecological collapse, wars, corruption, social unrest etc. What is AI supposed to align with, first of all, before we even figure out the "how" (as AI does not fall in the sphere of those relations described above that affect human alignment)?